Brain-based ranking of cognitive domains to predict schizophrenia

Abstract

Schizophrenia is a devastating brain disorder that disturbs sensory perception, motor action, and abstract thought. Its clinical phenotype implies dysfunction of various mental domains, which has motivated a series of theories regarding the underlying pathophysiology. Aiming at a predictive benchmark of a catalog of cognitive functions, we developed a data-driven machine-learning strategy and provide a proof of principle in a multisite clinical dataset (n = 324). Existing neuroscientific knowledge on diverse cognitive domains was first condensed into neurotopographical maps. We then examined how the ensuing meta-analytic cognitive priors can distinguish patients and controls using brain morphology and intrinsic functional connectivity. Some affected cognitive domains supported well-studied directions of research on auditory evaluation and social cognition. However, rarely suspected cognitive domains also emerged as disease relevant, including self-oriented processing of bodily sensations in gustation and pain. Such algorithmic charting of the cognitive landscape can be used to make targeted recommendations for future mental health research.

Keywords: BrainMap database; coordinate-based meta-analysis; ontology of the mind; pattern recognition; predictive analytics; statistical learning.

Figure 1

Overview of a taxonomy that compartmentalizes human cognition. (Left) Exhaustive set of mental operations used for brain‐driven ranking of altered cognitive concepts in schizophrenia. BrainMap defines two description systems: Mental domains (shown here) and experimental tasks (Figure S1). Note that the five top classes (action, etc.) were disregarded in the present study to avoid hierarchical dependence between the cognitive classes. This database offers results of almost a quarter of the published functional neuroimaging experiments carefully annotated with both taxonomies. (Right) A cognition‐topography map for each cognitive category was generated from the neuroimaging database. Four examples of cognitive meta‐priors are shown (z‐scored for display, only voxels with positive z‐scores shown) [Color figure can be viewed at http://wileyonlinelibrary.com]

Figure 2

Overview of our analysis workflow. Illustrates our approach to automatically rank a set of commonly studied cognitive processes for their predictive relevance in schizophrenia (SCZ). (a) We capitalized on two types of data resources. The BrainMap database provided existing neuroscience knowledge in form of robust neural activity changes reported in published neuroimaging experiments. Each experiment was labeled with the examined cognitive processes by means of a comprehensive cognitive taxonomy. Additionally, we built on structural and functional magnetic resonance imaging data (sMRI and fMRI) from a multisite dataset of patients with SCZ and healthy controls (HCs). (b) From the neuroimaging database, we quantitatively summarized the topography of consistently evoked neural activity changes associated with each cognitive domain (e.g., pain) into a “cognitive meta‐prior.” (c) The ensuing set of cognitive meta‐priors served as masks to extract cognitive domain‐specific information from structural and functional brain scans of SCZ patients and HCs. The data extraction followed nine complementary ways to aggregate neurobiological information, such as mining peak locations, local region, and integrative network characteristics. (d) To impartially rank the cognitive meta‐priors for their predictive value in SCZ, we used a two‐step approach. First, we built several base models to test each particular cognitive meta‐prior separately for its capability of telling patients and controls apart. Second, we combined the collection of all base model predictions into a higher level summary model encapsulating the entire cognitive taxonomy. The summary model put all meta‐priors on the same scale and could thus directly compare a variety of cognitive processes in their usefulness for detecting SCZ from brain scans. The resulting rankings of cognitive processes were averaged across the nine different neurobiological sampling strategies. (e) Finally, we validated the ability of the built predictive model to distinguish patients and controls in the future based on previously unseen, left‐out participants (10‐fold cross‐validation scheme) [Color figure can be viewed at http://wileyonlinelibrary.com]

Figure 3

Validation of our data‐analysis framework. The final predictive model classified healthy versus schizophrenic individuals statistically significantly better than a cognition‐naive null model in each of two taxonomies. We estimated the null distribution by selectively corrupting the participant pattern of cognitive indices while leaving other structure in the data intact. To ascertain that the final predictive model captured participant‐specific cognitive facets instead of confounding variables. Purple diamonds indicate the (out‐of‐sample) classification performance of the composite model based on (a) mental domains and (b) experimental tasks using combined structural (sMRI) and functional (fMRI) brain information. The dots show 1,000 model performances realized under the null hypothesis. The gray boxplots show bold lines for median (50th percentile), the lower and upper quartile (25th and 75th percentiles), and whiskers for the interquartile distance (25th–75th percentiles) besides the box. In each of nine ways to sample brain information, the composite model performed significantly better than the null model (p < .05 for mental domains and p < .01 for experimental tasks). If the individual combinations of cognitive expressions were not relevant, we would only observe our actually obtained prediction performance (purple diamond) in at most 50 out of 1,000 cases for mental domains and in at most 10 out of 1,000 cases for experimental tasks. The negative test implies that the successful individualized decisions of our predictive model can be ascribed to participant‐specific cognitive alterations rather than other characteristics of the participant sample [Color figure can be viewed at http://wileyonlinelibrary.com]

Figure 4

Quantified predictive value of mental domains in schizophrenia (SCZ). We systematically screened for dysregulated cognitive processes to facilitate the development of personalized diagnoses and new treatment strategies. Relative contribution of mental domains in disambiguating patients with SCZ and healthy controls. Thirty‐four mental domains ordered according to their average ability to forecast disease status. Joyplot shows weighted importance ranks for each domain (colored mountains). Red diamonds depict mean ranking position across brain sampling strategies (see Figure S3 for pipeline‐specific domain ranks). Certainty of discriminability position was assessed by estimating bootstrapped 95% population intervals (red lines). For instance, gustation was highly predictive across complementary approaches to sample neurobiological information, whereas the relevance of audition was more dependent on the sampling pipeline. Some intensively studied concepts of attention (e.g., Braff, 1993) and working memory (e.g., Forbes et al., 2009; Lee & Park, 2005) have been situated among the cognitive classes least predictive for SCZ. All results based on combined sMRI and fMRI data [Color figure can be viewed at http://wileyonlinelibrary.com]

Figure 5

Quantified predictive value of experimental tasks in schizophrenia (SCZ). We broadly screened for distinctive experimental paradigms to facilitate the development of personalized diagnoses and new treatment strategies. Relative contribution of experimental tasks in disambiguating patients with SCZ and healthy controls. Fifty experimental tasks were ordered according to their average ability to forecast disease status across brain sampling strategies (see Figure 4 for pipeline‐specific domain ranks). Joyplot shows weighted importance ranks for each domain (colored mountains). Red diamonds depict mean position in relevance. Certainty of discriminability position was assessed by estimating 95% population intervals (red lines). Precision estimates computed by repeatedly resampling participants with replacement (bootstrapping). All results based on combined sMRI and fMRI data [Color figure can be viewed at http://wileyonlinelibrary.com]

Figure 6

Domain–domain interactions in detecting schizophrenia (SCZ). Inspects in more detail potentially complicated relationships in the predictability of (a) mental domains and (b) experimental tasks after accounting for the influence of the remaining domains (instead of ignoring them). Partial dependence of SCZ predictability (z‐axis) on the joint distribution of two selected cognitive domains (x‐ and y‐axis) in predicting whose brain scans are from a SCZ patient (Principal component analysis [PCA] pipeline). Pairs of the top three cognitive processes are shown for both taxonomies (see Figure S5 for further examples). All results based on combined brain imaging types (sMRI and fMRI data) [Color figure can be viewed at http://wileyonlinelibrary.com]

Figure 7

Predictive relevance maps for schizophrenia (SCZ). Quantifies the average extent to which individual brain regions contributed to disease classification across (a) 34 mental domains and (b) 50 experimental tasks. Whole‐brain maps depict relative importance values of the cognitive meta‐priors across nine brain sampling strategies. In both taxonomies, nodes of the dorsal attention network (e.g., frontal eye field [FEF] and intraparietal sulcus [IPS]) and saliency network (e.g., anterior insula [AI] but not mid cingulate cortex [MCC]) as well as thalamus were highly pertinent in distinguishing patients from controls. Left and right temporo‐parietal junction (TPJ), however, emerged as discriminative in mental domains but not in experimental tasks. Both taxonomies provided largely overlapping but still distinct brain patterns underlying SCZ classification. This converging evidence across two independent cognitive taxonomies further strengthens the validity of our approach. Brain maps were smoothed (FWHM = 6mmm) and thresholded for display (see https://neurovault.org/collections/4074/ for unthresholded predictive brain maps). All results are based on combined sMRI and fMRI data [Color figure can be viewed at http://wileyonlinelibrary.com]